Automated Aesthetic Evaluation of Digital Product Designs Using Patch-Based Convolutional Neural Networks
Data and Metadata · 12.03.2026 · 5 · 790
DOI: 10.56294/dm2026790 ↗Kurz gefasst
Fachartikel in der Zeitschrift Data and Metadata, online erschienen am 12.03.2026. Verfasst von Pu Xia. Die Arbeit nennt 18 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Introduction
Evaluating the aesthetic quality of digital product designs is essential for improving user experience, engagement, and overall visual appeal. Traditional evaluation approaches, including manual scoring and basic convolutional neural network (CNN) models, often struggle to simultaneously capture fine-grained local features and global layout structures. As a result, aesthetic assessments are frequently inconsistent or suboptimal. Standard CNN-based methods tend to emphasize holistic image characteristics while overlooking localized design elements, whereas handcrafted texture descriptors lack flexibility and adaptability across diverse digital design compositions.
Objective
This research aims to develop an automated, accurate, and robust framework for evaluating the aesthetic quality of digital product designs by effectively capturing both local texture information and global layout patterns.
Method
A novel Patch-based Convolutional Neural Network with Barber Optimization (PCNN-BO) was proposed. The approach divides digital product images into uniform patches to enable patch-level feature extraction, allowing the model to capture fine-grained local textures alongside global spatial structures. Barber Optimization (BO) is employed to optimize patch selection and CNN hyperparameters, improving convergence speed, predictive accuracy, and the representation of nuanced aesthetic attributes. The model was trained and evaluated using the Digital Product Design Aesthetic Dataset from Kaggle, which includes annotated aesthetic scores and design attributes. Image preprocessing involved resizing, histogram equalization, and min–max normalization. Texture features were extracted using the Gray-Level Co-occurrence Matrix (GLCM), while the CNN learned combined local and global representations. The framework was implemented using Python, TensorFlow/Keras, OpenCV, and standard optimization libraries.
Results
The PCNN-BO model achieved superior performance, with an accuracy of 96,1% , precision of 96,1 %, recall of 95,4 %, F1-score of 95,7 %, and an MSE of 0,018.
Conclusions
PCNN-BO provides a robust, scalable, and precise solution for automated aesthetic evaluation of digital product designs, effectively outperforming conventional CNN-based and handcrafted feature methods.
Zitieren
APA
Xia, P. (2026). Automated Aesthetic Evaluation of Digital Product Designs Using Patch-Based Convolutional Neural Networks. Data and Metadata, 5, 790. https://doi.org/10.56294/dm2026790
Vancouver
Xia P. Automated Aesthetic Evaluation of Digital Product Designs Using Patch-Based Convolutional Neural Networks. Data and Metadata. 2026;5:790. doi:10.56294/dm2026790
BibTeX
@article{Xia2026,
author = {Xia, Pu},
title = {Automated Aesthetic Evaluation of Digital Product Designs Using Patch-Based Convolutional Neural Networks},
journal = {Data and Metadata},
year = {2026},
volume = {5},
pages = {790},
doi = {10.56294/dm2026790}
}
EUVIMED